Two questions prompted by the v6.1 release announcement, which is locked, so I'm posting here. Both concern the extended two-level parallelization, from a user who just hit its limit in v6.0.
1. Does -ni > 1 now work with ephwrite = .true.?
On v6.0 it does not. A job launched as
mpiexec -n 100 --ppn 100 epw.x -nk 10 -ni 10 -in epw.in
terminated at 0.07 s with EPW reporting that ephwrite = .true. is incompatible with image parallelization. Since EPW has no plane-wave parallelization, ranks must equal pools, and pools are bounded by the coarse k-point count — 64 in my case. That capped my anisotropic Eliashberg calculation at a single node with no way to extend it.
The release note says two-level parallelization "has been extended across EPW" and "supports parallelization of calculations involving both coarse and fine electron and phonon grids." Does that include the ephwrite path, or does the restriction remain?
2. Has v6.1 been built with the Intel oneAPI toolchain?
The GPU offloading note lists CUDA, OpenACC and OpenMP target offloading. I am on Aurora at ALCF (Intel Ponte Vecchio, oneAPI), currently running the QE 7.5 / EPW v6.0 CPU build. Kim, Liu, Tiwari, Margine and Giustino (arXiv:2603.10295) report v6.1 scaling on Aurora, so I assume the toolchain works — but I have not found installation guidance specific to oneAPI, and I would rather ask than guess before requesting a build from facility staff.
For context: I have just completed an independent reproduction of an anisotropic Migdal-Eliashberg study using v6.0 on Aurora (SrB3C3, Wang, Yan & Gao, PRB 103, 144515). Ten quantities agree with the published values, including the gap anisotropy ratio at 5 K to 0.05%. The whole campaign took 22 node-hours on single nodes. If measured Aurora timings or configuration details would be useful to the developers, I am glad to provide them.
Thank you, Rick Maghetti Resonance Materials LLC ALCF Director's Discretionary allocation 16463 (Aurora)
EPW v6.1 — does -ni > 1 now work with ephwrite, and has it been built with Intel oneAPI?
Moderator: stiwari
Re: EPW v6.1 — does -ni > 1 now work with ephwrite, and has it been built with Intel oneAPI?
Dear Rick,
1. Unfortunately, ephwrite is still not image parallelized. Apart from that, most of epw now supports two-level parallelization including coarse-grid q-parallelization over irreducible q-points.
2. Yes, EPW has been both built and benchmarked with intel oneAPI toolchain on Aurora.
Let us know if you have more questions regarding EPW v6.1.
Best regards,
Sabya.
1. Unfortunately, ephwrite is still not image parallelized. Apart from that, most of epw now supports two-level parallelization including coarse-grid q-parallelization over irreducible q-points.
2. Yes, EPW has been both built and benchmarked with intel oneAPI toolchain on Aurora.
Let us know if you have more questions regarding EPW v6.1.
Best regards,
Sabya.
-
RMaghetti
- Posts: 2
- Joined: Fri Jun 12, 2026 8:16 pm
- Affiliation: Independent Researcher, Resonance M
Re: EPW v6.1 — does -ni > 1 now work with ephwrite, and has it been built with Intel oneAPI?
Sabya,
Thank you — both answers are exactly what I needed, and the oneAPI confirmation resolves a question the Download and Install page left open.
One follow-up, since you are a co-author on arXiv:2603.10295. That paper reports 19–29× speedups on Aurora for v6.1. Given that ephwrite remains outside the image parallelization, do those figures apply to the anisotropic Eliashberg path, or are they measured on paths that do benefit? I ask because my workload is ephwrite plus laniso, and I would rather size expectations correctly before requesting a facility build.
On the offer of measured Aurora timings: I could not find any published benchmark for the anisotropic solve — the Benchmarks page carries the isotropic H₃S strong-scaling study only. Mine ran roughly 15–30 minutes per temperature at 3,471 irreducible Fermi-shell k-points on 64 ranks of a single node, with memory at 15.2 Gb per pool, about 94% of the node. If a measured figure for that path would be useful to the developers, I am glad to write it up properly with the full configuration.
Thank you again,
Rick Maghetti
Thank you — both answers are exactly what I needed, and the oneAPI confirmation resolves a question the Download and Install page left open.
One follow-up, since you are a co-author on arXiv:2603.10295. That paper reports 19–29× speedups on Aurora for v6.1. Given that ephwrite remains outside the image parallelization, do those figures apply to the anisotropic Eliashberg path, or are they measured on paths that do benefit? I ask because my workload is ephwrite plus laniso, and I would rather size expectations correctly before requesting a facility build.
On the offer of measured Aurora timings: I could not find any published benchmark for the anisotropic solve — the Benchmarks page carries the isotropic H₃S strong-scaling study only. Mine ran roughly 15–30 minutes per temperature at 3,471 irreducible Fermi-shell k-points on 64 ranks of a single node, with memory at 15.2 Gb per pool, about 94% of the node. If a measured figure for that path would be useful to the developers, I am glad to write it up properly with the full configuration.
Thank you again,
Rick Maghetti
Re: EPW v6.1 — does -ni > 1 now work with ephwrite, and has it been built with Intel oneAPI?
Hi Rick,
The reported speedups are for electron-phonon matrix element interpolation and downstream calculations that benefit. That includes almost everything at this point except "ephwrite" where more than 1 image is not allowed. However, electron-phonon interpolation should benefit from GPU acceleration.
For Anisotropic Eliashberg calculations, we are working on image-parallelization and should be available in the develop branch of EPW (https://gitlab.com/epw/q-e.git) in near future.
Let me know if you have any further questions.
Best,
Sabya.
The reported speedups are for electron-phonon matrix element interpolation and downstream calculations that benefit. That includes almost everything at this point except "ephwrite" where more than 1 image is not allowed. However, electron-phonon interpolation should benefit from GPU acceleration.
For Anisotropic Eliashberg calculations, we are working on image-parallelization and should be available in the develop branch of EPW (https://gitlab.com/epw/q-e.git) in near future.
Let me know if you have any further questions.
Best,
Sabya.